artificial agents without ontological access to reality
TRANSCRIPT
Ar#ficial Agents Without Ontological Access to Reality
Olivier Georgeon h8p://liris.cnrs.fr/ideal/mooc
h8p://liris.cnrs.fr/ideal/mooc 1/31
Defini#ons
• Ontology: “Onto” (to be) + “Logos” (discourse): – Discourse on what “is”.
• Agent without ontological access to reality: – Agent that don’t have access to what “is” in reality. – Agent whose input data is NOT a representa5on of reality.
• We do not consider input data as the agent’s percep#on… • … then input data should be considered as what?
h8p://liris.cnrs.fr/ideal/mooc 2/31
Mainstream philosophy/epistemology • Philosophy
– Kant : Cogni#ve agent’s don’t have access to reality “as such” (noumenal reality).
• Psychology – Findlay & Gilchrist (2003). Ac#ve Vision.
• Cogni#ve Science. – “Percep#on and ac#on arise together, dialec#cally forming each other” (Clancey, 1992,
p5).
• Construc#vist epistemology – Piaget: Percep#on and ac#on are inseparable (sensorimotor schemes).
• Even quantum physics? – Predicts results of experiments without assuming an objec#ve state of reality
(talking about the state of Schrödinger’s cat makes no sense)
h8p://liris.cnrs.fr/ideal/mooc 3/31
Scope of this presenta#on
• Philosophical prerequisite: – Cogni5ve agents have no access to reality “as such”
• Our claim: – Most BICAs and machine learning algorithms have not yet acknowledged this philosophy!
• Content of the presenta#on: – How can we implement this philosophy into BICAs? – What will we gain and loose from doing so?
h8p://liris.cnrs.fr/ideal/mooc 4/31
The interac#on cycle
Agent
Environment
Input data
Output data
The agent interacts with the environment by receiving input data and sending output data. When does the interac#on cycle begin and end?
h8p://liris.cnrs.fr/ideal/mooc 5/31
Symbolic modeling
Agent Seman&c rules
Reality
Symbol Action
The agent receives a symbol that matches seman5c rules. There is a predefined “discourse on what is” (the set of symbols and seman#c rules) and the agent has access to it. The agent is a passive observer of reality. The cycle begins with the agent receiving input data and ends with the agent sending output data.
h8p://liris.cnrs.fr/ideal/mooc 6/31
Reinforcement learning
Agent
Reality st ∈ S
Action
at ∈ A
Observation ot = f (st) ∈ O Reward rt = r (st) ∈ ℝ
There is a predefined “discourse on what is” (the set S). Most reinforcement learning algorithms assume that the observa#on represents the state of reality (par#ally and with noise). The agent is a passive observer of reality. The cycle begins with the agent receiving input data and ends with the agent sending output data.
h8p://liris.cnrs.fr/ideal/mooc 7/31
Experiment / Result cycle
Agent
Experiment Result rt ∈ R xt ∈ X
Reality
The cycle begins with the agent sending output data and ends with the agent receiving input data. The agent is an ac#ve observer of reality (embodiment paradigm).
h8p://liris.cnrs.fr/ideal/mooc
We can’t assume that input data represent the state of reality: it may not! Most BICAs and Machine learning algorithms fail to generate interes#ng behaviors.
In a given state of reality, rt varies depending on xt.
8/31
Comparison
h8p://liris.cnrs.fr/ideal/mooc
Agent
Reality
Agent
Reality
a) Tradi#onal model b) Embodied model
a) and b) are mathema#cally equivalent but : -‐ a) highlights the common assump#on that input data represents reality. -‐ b) highlights that this assump#on may be wrong.
9/31
it it ot ot+1
Agents Without Ontological Access (AWOAs) are “indie” computer science
• Ar#ficial Intelligence (Russell & Norvig 2010, p. iv). – ”The problem of AI is to build agents that receive percepts from the environment
and perform ac5ons” – The problem of AI is to build agents that receive data (that may not be percepts)
from the environment and make decisions (that may not be ac5ons). • Reinforcement learning: (Su8on & Barto 1998, p. 4).
– “Clearly, such an agent must be able to sense the state of the environment to some extent and must be able to take ac#ons that affect the state. The agent also must have goals rela5ng to the state of the environment.”
– The agent must have preferences (drives) that may not relate to the state of the environment “as such”.
• AWOAs relate to other “indie” approaches to AI: – Enac#on, embodied cogni#on, developmental learning, mul# agent systems, etc.
• AWOAs differ from tradi#onal AI by design rather than by technique – All techniques can be used in both ways (rule based systems, connec#onist, mul#-‐
agent systems, reinforcement learning techniques, etc.)
h8p://liris.cnrs.fr/ideal/mooc 10/31
Example
(-‐3) (-‐3) (-‐1)
(-‐1)
(5)
(-‐10)
Set E of 6 experiments:
Set R of 2 results:
Set I = E x R of 12 interac#ons (with valence):
(-‐1)
(-‐1)
(-‐1)
(-‐1)
0 or 1
h8p://liris.cnrs.fr/ideal/mooc
The Agent / Environment coupling affords hierarchical regulari#es of interac#on, e.g.,
-‐ Amer , experiment results more likely in than in .
-‐ Amer , sequence can omen be enacted. -‐ Amer , sequence can omen be enacted.
11/31
The "Little loop problem"
http://liris.cnrs.fr/ideal/mooc
Bump: Touch:
Move Forward or bump (5) (-‐10) Turn lem / right (-‐3) Feel right/ front / lem (-‐1)
12/31
4. Afford
Time
6. Choose Decision Time
3. Ac#vate 5. Propose
7. Enact
1.
h8p://liris.cnrs.fr/ideal/mooc
Hierarchical bo8om-‐up sequence learning
13/31
Experiment Result
c) Experiment/Result Model
r ∈ R x ∈ X Agent
Intended interaction
Enacted interaction
i = 〈x,r〉 ∈ X×R
d) Interactional Model
Reality
Agent
Reality
e = 〈x,r’〉 ∈ X×R
Interac#onal model
Embodied models: the agent must use the active capacity of its body to make experiments in order to learn about reality.
h8p://liris.cnrs.fr/ideal/mooc 14/31
Agent
Environment
Environment “known” at time td
ecd ∈ Cd icd ∈ Cd
ep1 ip1 ipj ∈ I epj ∈ I
Decisional mechanism
Recursive learning and self-‐programming
h8p://liris.cnrs.fr/ideal/mooc 15/31
Ac#vity analysis
10 20 30 40 50 60 70 80 90 100
100 110 120 130 140 150 160 170 180 190 200
200 210 220 230 240 250 260 270 280 290 300
300 310 320 330 340 350 360 370 380 390 400
touch front – move forward (step 74)
10 20 30 40 50 60 70 80 90 100
100 110 120 130 140 150 160 170 180 190 200
200 210 220 230 240 250 260 270 280 290 300
300 310 320 330 340 350 360 370 380 390 400
touch lem – turn lem – move forward (Step 186)
10 20 30 40 50 60 70 80 90 100
100 110 120 130 140 150 160 170 180 190 200
200 210 220 230 240 250 260 270 280 290 300
300 310 320 330 340 350 360 370 380 390 400
h8p://liris.cnrs.fr/ideal/mooc 16/31
e-‐puck robot (it resists to noise!)
h8p://liris.cnrs.fr/ideal/mooc 17/31
It allows training
h8p://liris.cnrs.fr/ideal/mooc 18/31
Rudimentary distal perception
!"#$%!%
!"#$%&%
!"#$%'%
()*+$,%*#-*."/%0#,1%
Detects rela#ve displacement of objects and approximate direc#on within 180° span (area A, B, or C). “Likes” rapprochement. “Dislikes” disappearance.
h8p://liris.cnrs.fr/ideal/mooc 19/31
Self-‐programming
h8p://liris.cnrs.fr/ideal/mooc 20/31
No free lunch for machine learning • It does not violate the “no free lunch theorem”
– Wolpert, D.H., & Macready, W.G. (1997) • What we loose:
– It does not learn to reach predefined goal states. • e.g., win at chess.
• What we gain – It learns hierarchical sa#sfying habitudes much faster. – Prac#cal applica#ons when we need systems to learn habitudes:
• e.g, home automa#on, somware adapta#on, end-‐user programming… – Robots that interact with the real world (without predefined model)
– Theore#cal applica#ons • It opens the way to higher-‐level cogni#on (if we trust Kant, Piaget, etc.)
h8p://liris.cnrs.fr/ideal/mooc 21/31
AImergence
h8p://www.oliviergeorgeon.com/aimergence
h8p://liris.cnrs.fr/ideal/mooc 22/31
Non-‐Markov Reinforcement Learning
Stage 1 Stage 2 Stage 3
40 200 240 320 480 520 640 800 840-15
-12.5
-10
-7.5
-5
-2.5
0
2.5
o1.a1… on.an.on+1, |o1… on+1∈O and a1… an∈A.
h8p://liris.cnrs.fr/ideal/mooc 23/31
Blue phenomenon
White phenomenon
Level 3
Unknown
123456
1
? ?
10
? ? ? ?
20
? ? ? ? ? ? ?
30
?
40 50 60 70
Time h8p://liris.cnrs.fr/ideal/mooc 24/31
Agent
Intended Interaction
Enacted interaction
i = 〈x,r〉 ∈ X×R
d) Interactional model
Reality
Agent
Intended experience
Enacted experience
e ∈ E i ∈ E
e) Experiential model
Reality
e = 〈x,r’〉 ∈ X×R
Experien#al model
It’s a radical inversion of our viewpoint on artificial agent: We focus on the agent’s stream of phenomenological experience
123456
1
? ?
10
? ? ? ?
20
? ? ? ? ? ? ?
30
?
40 50 60 70
h8p://liris.cnrs.fr/ideal/mooc 25/31
Agent Intended experiences I ⊂ Σ
Enacted experiences
E ⊂ Σ
Reality
Spatial displacement
𝜏
Spatial coupling
Σ : Experiences with spatial attributes
h8p://liris.cnrs.fr/ideal/mooc 26/31
Interac#on Timeline
Egocentric Spa#al Memory
Hierarchical Sequen#al System
Behavior Selec#on
Intend
Prop
ose
Propose
Learn / Track
Ontology
Evoke
Construct
Enact
AGENT
h8p://liris.cnrs.fr/ideal/mooc 27/31
Spatial Little Loop Problem
http://liris.cnrs.fr/ideal/mooc 28/31
Dynamic environment
h8p://liris.cnrs.fr/ideal/mooc 29/31
Robo#cs research
h8p://liris.cnrs.fr/ideal/mooc
Bumper tac#le sensor
Panoramic camera
Ground op#c sensor
h8p://liris.cnrs.fr/simon.gay/index.php?page=eirl&lang=en 30/31
Conclusion: a research approach • Theory of ar#ficial Agents Without Ontological Access to reality
(AWOA) is under development. • We design embodied models that focus on the agent’s stream of
phenomenological experience. • We validate the agents through behavioral analysis rather then
through performance measures. • Create animal-‐level intelligence before human-‐level intelligence.
– “Animal-‐level Turing test” based on behavioral analysis?
• We (as a community) must define criteria of intelligent behavior. • Incremental approach: imagine increasingly difficult experiments and
design smarter agents in parallel. – (aimergence game).
h8p://liris.cnrs.fr/ideal/mooc 31/31